A robust and optimal matching-score network is presented. The algorith
m is based on a steepest descent training procedure. The net is capabl
e of learning new patterns without recomputing the entire interconnect
ion weight matrices. Simulation results are given to illustrate the st
rengths of the proposed design method.